AMD x Higgsfield DoP x TensorWave

Apr 2, 2025

At Higgsfield AI, we’re building generative video products tailored for creative professionals — and high-performance, efficient infrastructure is foundational to what we do. As part of that mission, we developed Higgsfield DoP I2V-01-preview, our proprietary Image-to-Video (I2V) model designed for high-quality video generation from images. In collaboration with TensorWave, a cloud provider offering AMD-based compute instances, we benchmarked and profiled our model on AMD Instinct™ MI300X GPUs, evaluating its performance across inference workloads and real-world deployment scenarios.

Higgsfield DoP I2V-01-preview brings cinematic structure and control into generative video through a novel architecture that blends diffusion models with reinforcement learning. Rather than simply denoising frames, the model is trained to understand and direct motion, lighting, lensing, and spatial composition — capturing the grammar of cinematography. Inspired by how reinforcement learning has been used to teach large language models reasoning skills, we applied RL after diffusion to instill intent and coherence in generated sequences. The result is a system capable of producing expressive, controllable, and high-fidelity video — powered by a robust infrastructure stack purpose-built for professional creative workflows.

Inference with Tensorwave on AMD Instinct™ GPUs 

TensorWave’s AMD-based infrastructure, built on next-generation accelerators, provided a scalable and memory-optimized environment ideally suited for our inference workloads. With pre-configured PyTorch and ROCm™environments, we were able to run inference out of the box — no custom setup or engineering effort required. This streamlined deployment experience allowed us to focus entirely on validating the stability and performance of our model on AMD Instinct™ MI300X GPUs.

Profiling I2V Inference: Where the Time Goes

Before diving into performance benchmarks, we first validated stability and correctness.

We wanted to ensure that running our model on AMD wouldn't introduce:

  • Unexpected slowdowns

  • Kernel mismatches or fallback ops

  • Memory leaks or crashes

  • Subtle bugs in attention mechanisms

None of these issues occurred. Everything ran smoothly from the start — no workarounds, no surprises. With stability and correctness fully validated, we proceeded to assess runtime performance under production-like conditions.

Our generation speed on AMD MI300X outperformed the same workload running on Nvidia H100 SXM. In our internal benchmarks, generating videos at 1280x720 (720p) resolution with 20 inference steps was consistently faster on AMD. Even more notably, when scaling to 1080p resolution, the H100 frequently ran into out-of-memory (OOM) issues — while MI300X handled the workload out of the box, thanks to its significantly larger memory capacity. This demonstrates that ROCm isn’t just a functional alternative — it delivers real performance gains. With solid kernel support and optimized transformer implementations, AMD hardware proves fully capable of meeting — and even exceeding — the performance of leading inference platforms for generative video.

AMD and Higgsfield DoP: Enabling Scalable, Proprietary Video Generation

As demand grows for high-performance generative video tools, AMD continues to support developers and creators through scalable, efficient hardware. At Higgsfield AI, we’ve developed our proprietary DoP I2V-01-preview model to power next-generation, creator-focused video generation products. AMD Instinct™ MI300X GPUs have proven to be a robust foundation for running our model efficiently right out of the box. With broad GPU availability and an open software stack via ROCm™, AMD enables fast, flexible deployment

of complex AI workloads. We’re excited to collaborate with AMD and the ecosystem to push the boundaries of generative video and build powerful creative tools for the future.

Additional Resources: